Comment by nothrabannosir

I think the "floating point are bad for storing currencies" is one of the most common misconception about floating point.

Most people don't realize that the IEEE-754 single precision floating point represent real numbers with 9 decimal digits (or 23 binary digits). The double, on the other hand, represents the real numbers with 17 decimal digits.

This means that the double error UPPER BOUND is (0.00000000000000001)/2 per operation. But in reality the error is lower because of the rounding operations.

Also, it is posssible to extend the range using denormals, but most (all?) compilers disable them when compiling with anything other than O0 to avoid performance degradation.

The overheads associate with dealing with non-float types for most applications might not be worth it the cost and risk. If course, if the language are working with provides a currency type, go for it. But if doesn't , there is no need to worry.

I agree with your overall point: it most likely does not matter when the values are close enough. However :)

There can be two companies with 100M market cap. Corp A has issued 10M shares @ 10 each, Corp B has 10B shares priced at 0.01

A +/-0.001 change in Corp A share price is just 0.01% and moves the market cap by +/-10k, so probably nothing significant. The same nominal change in Corp B amounts to 10%, or +/- 10M in the company value, which is quite a big deal.

Also I think there may be some money to be made in changes at the 7th decimal place with large enough volume of high frequency transactions.

And that roughly captures the spot where I was seeing doubles used.

Yes, they could have used fixed point. I am guessing that what happened is that someone who had thought way more deeply about this than I ever needed to (I worked on the accounting side, where, yep, we always used decimals) either determined that, where the modeling was concerned, floating point errors were not worth worrying about, or estimated that the expected cost to the company stemming from bugs due to to fixed point math being easier to goof up on would have been smaller than the expected cost to the company due to floating point error.

Financial models are predictive, they don't have to be accurate to a penny, right? Unlike processing actual money people own.

(I do some work with predictive simulations about money, but outside finance, and there we care that the result has accurate order of magnitude. Floats were used extensively in the project; I actually upgraded them to doubles for the sake of handling larger order of magnitude spans.)

I stand corrected, thanks for this example.

Amazon's EC2 hourly prices are rounded to mils ($0.011/hour).

https://aws.amazon.com/emr/pricing/

Azure has some hourly prices with ten-thousandths of a cent ($0.0102/hour):

https://azure.microsoft.com/en-ca/pricing/details/virtual-ma...

Microsoft should use gas station 9/10 pricing conventions to just barely undercut Amazon's lowest price $0.011 with $0.0109.

https://www.marketplace.org/2018/10/11/why-do-gas-prices-end...

>“They found out that if you priced your gas 1/10 of a cent below a break point, let’s say 40 cents a gallon, ‘.399’ just looked to the public like 39 cents…”

> Performing arithmetic operations against money in floating point is the dangerous part, as error can accumulate beyond an atomic unit.

A good example of this is trying to compute the sales tax on $21.15 given a tax rate of 10%. The exact answer would be $2.115, which should round to $2.12.

IEEE 64-bit floating point gives 2.1149999999999998, which is hard to get to round to 2.12 without breaking a bunch of other cases.

Here are three functions that try to compute tax in cents given an amount and a rate, in ways that seem quite plausible:

  def tax_f1(amt, rate):
    tax = round(amt * rate,2)
    return round(tax * 100)
  
  def tax_f2(amt, rate):
    return round(amt*rate*100)
  
  def tax_f3(amt, rate):
    return round(amt*rate*100+.5)

On these four problems:

   1% of $21.50
   3% of $21.50
   6% of $21.50
  10% of $21.15

the right answers are 22, 65, 129, and 212. Here are what those give:

  tax_f1:  21  65 129 211
  tax_f2:  22  64 129 211
  tax_f3:  22  65 130 212

Note that none of the get all four right.

I did some exhaustive testing and determined that storing a money amount in floating point is fine. Just convert to integer cents for computation. Even though the floating point representation in dollars is not exact, it is always close enough that multiplying by 100 and rounding works.

Similar for tax rates. Storing in floating point is fine, but convert to an integer by multiplying by an appropriate power of 10 first. In all the jurisdictions I have to deal with, tax rate x 10000 will always be an integer so I use that.

Give amt and rate, where amt is the integer cents and rate is the underlying rate x 10000, this works to get the tax in cents:

  def tax(amt, rate):
    tax = (amt * rate + 5000)//10000
    return tax

I'm not fully convinced that you cannot do all the calculations in floating point, but I am convinced that I can't figure it out.

> Storing money in floating point is fine. Just round to the nearest atomic unit when displaying.

Well, it's not just a display issue. In accounting, associativity and commutativity are important. People do care that `a + b + c - a == c + b` should evaluate to “true”.

There's very little point in storing money in floats if you're not going to do arithmetic in floats; about the only use case I can think of is JavaScript and JSON APIs.